Raw Acceleration Data Research Articles

Relationship Between Parent and Child Physical Activity Using Novel Acceleration Metrics

ABSTRACT Background: Raw acceleration data lend themselves to innovative metrics in which comparisons can be made across populations. This study examined whether the intensity gradient (IG) or average acceleration (AvAcc) was associated with body mass index (BMI) in children and adults and explored parent–child associations between time spent in physical activity intensities and BMI. Methods: Parent–child dyads (n = 90) wore a wrist-mounted ActiGraph GT3X+ monitor for ≥ 3 valid days (i.e., ≥ 16 hrs each day). Participants’ BMI was calculated. Results: Girls’ BMI-z scores were positively associated with parent BMI and inactive time. Parent BMI was negatively associated with boys’ inactive time. Parental inactive time predicted girls’ BMI-z scores. The metrics were not associated with parent BMI, and no independent effects were observed between the metrics. Associations between boys’ IG and BMI z-score were observed. No independent effects were observed between the metrics and no significant associations were observed between boys’ BMI z-score and AvAcc. Girls’ AvAcc and IG were not associated with BMI z-score and no independent effects were observed between the metrics. Conclusion: Findings support previous research which suggests that parent’s activity does not influence their child’s activity.

Evaluating the Performance of Sensor-based Bout Detection Algorithms: The Transition Pairing Method.

To present and demonstrate the Transition Pairing Method (TPM), a new method for evaluating the performance of bout detection algorithms. The TPM compares predicted transitions to a criterion measure in terms of number and timing. A true positive is defined as a predicted transition that corresponds with one criterion transition in a mutually exclusive pair. The pairs are established using an extended Gale-Shapley algorithm, and the user specifies a maximum allowable within-pair time lag, above which pairs cannot be formed. Unpaired predictions and criteria are false positives and false negatives, respectively. The demonstration used raw acceleration data from 88 youth who wore ActiGraph GT9X monitors (right hip and non-dominant wrist) during simulated free-living. Youth Sojourn bout detection algorithms were applied (one for each attachment site), and the TPM was used to compare predicted bout transitions to the criterion measure (direct observation). Performance metrics were calculated for each participant, and hip-versus-wrist means were compared using paired T-tests (α = 0.05). When the maximum allowable lag was 1-s, both algorithms had recall <20% (2.4% difference from one another, p<0.01) and precision <10% (1.4% difference from one another, p<0.001). That is, >80% of criterion transitions were undetected, and >90% of predicted transitions were false positives. The TPM improves on conventional analyses by providing specific information about bout detection in a standardized way that applies to any bout detection algorithm.

A hybrid method coupling empirical mode decomposition and a long short-term memory network to predict missing measured signal data of SHM systems

Missing data, especially a block of missing data, inevitably occur in structural health monitoring systems. Because of their severe negative effects, many methods that use measured data to infer missing data have been proposed in previous research to solve the problem. However, capturing complex correlations from raw measured signal data remains a challenge. In this study, empirical mode decomposition is combined with a long short-term memory deep learning network for the recovery of the measured signal data. The proposed hybrid method converts the missing data imputation task as a time series prediction task, which is then solved by a “divide and conquer” strategy. The core concept of this strategy is the prediction of the subsequences of the raw measured signal data, which are decomposed by empirical mode decomposition rather than directly predicted, as the decomposition can assist in the modeling of the irregular periodic changes of the measured signal data. In addition, the long short-term memory network in the hybrid model can remember more long-range correlations of subsequences than can the traditional artificial neural network. Three widely used prediction models, namely, the autoregressive integrated moving average, support vector regression, and artificial neural network models, are also implemented as benchmark models. Raw acceleration data collected from a cable-stayed bridge are used to evaluate the performance of the proposed method for missing measured signal data imputation. The recovery results of the measured signal data demonstrate that the proposed hybrid method exhibits excellent performance from two perspectives. First, the decomposition by empirical mode decomposition can improve the accuracy of the core long short-term memory prediction model. Second, the long short-term memory model outperforms other benchmark models because it can fit more microscopic changes of measured values. The experiments conducted in this study also suggest that the change patterns of raw measured signal data are complex, and it is therefore important to extract the features before modeling.

Dynamic Segmentation Of Youth Accelerometer Data By Sojourn And Change Point Detection Methods

Dynamic segmentation algorithms are used to find activity transitions in accelerometer data. Youth Sojourn models use a crude algorithm, which may be improved by instead using a change point detection (CPD) algorithm. Pruned exact linear time (PELT) is a CPD algorithm that finds transitions by minimizing a cost function while iterating over the data and pruning out inviable transition points. PURPOSE: To compare the performance of youth Sojourn and PELT. METHODS: Raw acceleration data (hip-worn ActiGraph GT9X) from 86 youth (age 6-18 yrs; 48% male; 16% overweight/obese) were processed using Sojourn and PELT. Participants performed two semi-structured activity routines on separate days, with each visit lasting approximately 2-2.5 h. A total of 16 activities (eight each day) were performed, twice each, and the study protocol was designed to promote variability in the ordering and duration of activities. Throughout each trial, direct observation was performed using focal sampling, which served as a criterion measure of when activity transitions occurred. Sojourn and PELT were compared to the criterion using the transition pairing method, with a maximum of 5-s lag time allowed for a prediction to be considered a true positive. Performance metrics were recall, precision, and root mean squared error (RMSE). The metrics were calculated for each participant (both visits combined), after which paired t-tests were used to compare Sojourn-vs-PELT means for each metric. RESULTS: Values are mean ± SD. Recall was similar for Sojourn (49.6% ± 9.0%) and PELT (51.5% ± 9.2%, p = 0.15), and the same was true for RMSE (2.9 ± 0.3 s for Sojourn, versus 3.1 ± 0.4 s for PELT, p < 0.001). However, precision for Sojourn (21.7% ± 4.9%) was substantially lower than for PELT (38.7% ± 11.0%, p < 0.001). CONCLUSION: Youth Sojourn models may benefit from replacing their current segmentation algorithms with CPD algorithms like PELT. Thus, CPD warrants further investigation. Supported by NIH R01HD083431

Classification Accuracy Of Wrist-worn Physical Activity Monitors Relative To Free-living Heart Rate

Wrist-worn accelerometers have replaced hip-worn devices as the wear-site of choice when measuring physical activity (PA) in many large-scale studies. Data suggesting superior compliance with study protocols has largely driven this transition due to the potential for a more accurate view of habitual PA. Similarly, activity classification utilizing raw acceleration data has gained popularity relative to epoch-based activity count methods, owing to open source analytical packages, higher classification accuracy, and the potential for greater comparability among devices. As activity classification methods for wrist-worn accelerometer data are derived from PA performed in controlled settings, their accuracy in quantifying free-living PA is unknown. PURPOSE: The purpose of this study was to examine the classification accuracy of common PA quantification methods against a free-living, participant-specific intensity classification, heart rate reserve (HRR). METHODS: Healthy young adults (n=33; 18.6 ± 0.7 years, 69.6% female) wore a triaxial accelerometer on their non-dominant wrist and a heart rate monitor around their chest for 24 hours. Free-living intensity was quantified using traditional HRR ranges (e.g. MVPA ≥40%), calculated using resting heart rate during sleep and age-predicted maximum heart rate. Two commonly used data classification methods were applied, 1) Euclidean norm minus one (ENMO) values calculated from raw triaxial data using milligravitational (mg) cut points of light <100.6, moderate 100.6 - <428.8, vigorous ≥428.8, and MVPA ≥100.6, and 2) activity counts across 1-, 15-, and 60-second epochs with count per minute (cpm) cut points of light 1514 - <2199, moderate 2199 - <4712, vigorous ≥4721, and MVPA ≥2199. RESULTS: ENMO-based classification underestimated average MVPA by 31.9 ± 106.6% (19.4 ± 105.3 mins). In contrast, activity count-based classification overestimated average MVPA by 368.4 ± 396.2% (178.3 ± 105.1 mins), 708.9 ± 914.7% (271.0 ± 113.9 mins), and 798.9 ± 1170.7% (280.3 ± 131.7 mins), for 1-, 15-, and 60-second epochs, respectively. CONCLUSIONS: Data processing utilizing raw triaxial data and ENMO appears to offer the most accurate PA intensity classification. However, all methods substantially misclassify activity relative to free-living HRR.

Prediction of bite number and herbage intake by an accelerometer-based system in dairy sheep exposed to different forages during short-term grazing tests

The accurate estimation of herbage intake is a key to adequately feed grazing ruminants. Ten dairy Sarda sheep fitted with a halter equipped with an accelerometer (BEHARUM device) were allowed to graze micro-swards of Italian ryegrass (Lolium multiflorum L.), alfalfa (Medicago sativa L.), oat (Avena sativa L.), chicory (Cichorium intibus L.) and a mixture (Italian ryegrass and alfalfa) for six minutes. Accelerometer data and video recordings of behaviour were collected simultaneously. The raw acceleration data were processed to calculate 15 variables: sum, mean, variance and inverse coefficient of variation (ICV, mean/standard deviation) for the X, Y and Z axes and the resultant. A database was created that included the acceleration variables and herbage intake (FMI, DMI, g), intake rate (FMIR, DMIR, g/minute), bite mass (FMBM, DMBM, g) either on fresh (FM) or dry matter basis (DM) and the logarithm of number of bites (LB) and bite rate (LBR) measured during the tests. Partial least square regression analysis (PLSR) was used to verify if acceleration variables could be used as predictors of behavioural traits. The precision and accuracy of PLSR were evaluated implementing the Model Evaluation System, in which predicted values were regressed against observed ones, based on r2 and Dent & Blackie test. The PLSR showed an overall good accuracy (Dent & Blackie test P = ns) and was proven precise for the estimation of LB (r2 = 0.87), LBR (r2 = 0.86), DMI and DMIR (r2 = 0.71). To conclude, BEHARUM can accurately estimate with high to moderate precision number of bites and herbage intake of sheep short term grazing Mediterranean forages.

Twenty four-hour activity cycle in older adults using wrist-worn accelerometers: The seniors-ENRICA-2 study.

This study aimed: (a) to provide a detailed description of sleep, sedentary behavior (SED), light physical activity (LPA), and moderate-to-vigorous physical activity (MVPA) over the complete 24-hours period using raw acceleration data in older adults; and (b) to examine the differences in the 24-hours activity cycle by sex, age, education, and body mass index (BMI). Population-based cohort comprising 3273 community-dwelling individuals (1739 women), aged 71.8±4.5years, participating in the Seniors-ENRICA-2 study. Participants wore a wrist-worn ActiGraph GT9X accelerometer for 7 consecutive days, and the raw signal was processed using the R-package GGIR. Participants reached 21.5mg as mean acceleration over the whole day; 32.3% (7.7h/d) of time was classified as sleep, 53.2% (12.7h/d) as SED, 10.4% (148.6min/d) as LPA, and 4.1% (59.0min/d) as MVPA. No marked differences were found in sleep-related variables between socio-demographic and BMI groups. However, women showed higher LPA but lower SED and MVPA than men. Moreover, SED increased whereas LPA and MVPA decreased with age. Participants with obesity (BMI≥30kg/m2 ) accumulated more SED and less LPA and MVPA than those without obesity. As expected, adherence to physical activity recommendations varied widely (9.2%-76.6%) depending on the criterion of MVPA accumulation. Objective assessment of the 24-hour activity cycle provides extensive characterization of daily activities distribution in older adults and may inform health-promotion interventions in this population. Women, the oldest old, and those with obesity offer relevant targets of strategies to improve lifestyle patterns.

Mobility of dry granular flows of varying collisional activity quantified by smart rock sensors

Highly instrumented particles (i.e., “smart rocks”) were included in monodisperse dry granular landslide experiments to quantify the collisional nature of such flows and to investigate the influence of collisional flow on the mobility of landslides. The total number of particles comprising a constant source volume of 0.4 m3 was varied by filling the volume with monodisperse particles of nominal diameters of 3, 6, 13 or 25 mm. Successively raising the total particle count resulted in flows that were increasingly thick relative to the respective particle size. Raw resultant acceleration data from the embedded smart rock sensors indicate that for each increase in grain size, there were increases in both the magnitude and frequency of particle collisions. Light detection and ranging (LiDAR)-generated point clouds of the landslide deposits indicated that increases in mobility and spreading, compared using differences in travel angle, were directly proportional to increases in collisional activity. By changing the size of the landslide particles from 3 to 25 mm, the travel angle at the gravity centre (αg) was observed to decrease from 27.8° to 25.3° (Δαg = −9.0%) and the Fahrböschung angle (α) was observed to decrease from 25.0° to 21.4° (Δα = −14.4%).

WaveNet: learning to predict wave height and period from accelerometer data using convolutional neural network

Inertial sensors carried by buoys, such as accelerometers, are widely used in wave characteristics measurement. Traditional methods usually employ numerical integration on the accelerate data for wave height, where the “drifting” errors are intractable. In this paper we propose a novel method to predict wave height and period using machine learning approach, specially a convolutional neural network. The end-to-end 1D convolutional neural network named WaveNet predicts wave height and period from the raw acceleration data directly. We designed a simple device to simulate the motion of the buoy in the wave, and used it to collect data for training and testing our model. The results of the proposed method were compared with traditional numerical integration method and found that the proposed model outperforms existing method in outputting more accurate wave height and period.

The peak player load™ of state-level netball matches

ObjectivesTo investigate the peak accelerometer-derived intensity of state-level netball matches and compare differences between positional groupings. Findings will provide guidance for sport science professionals on how to best replicate the most intense passages of play in training settings. DesignLongitudinal (one season). MethodTwenty-eight netball athletes across three teams from the same club wore an accelerometer (S5 Optimeye, Catapult sports) for all matches, in one season. Raw acceleration data were downloaded and converted into a vector magnitude (Player Load™) we then quantified the peak intensity over 30-s and one to ten-minute time periods. Positional groupings were created based on the number of thirds on a netball court that a particular position can enter, as this was deemed more appropriate for the current study than the traditional combinations based on tactical requirements. A linear mixed-model with fixed and random effects was utilised along with magnitude-based inferences to determine meaningful differences with 90 % confidence limits (CL). ResultsAcross all time periods post 30-s, only one comparison was not meaningfully different i.e. three-thirds v two-thirds at the one-minute timepoint (effect size: 0.27, CL −0.05 to 0.60). ConclusionsFindings justify that netball athletes, depending on positional group defined by this study, should train at different intensities dependent on a specified duration. Conditioning professionals and coaches should design training drills that best replicate the peak intensity of match play. This may improve an athlete’s physical performance capacity during highly exertive periods of competition, which regularly occur at critical moments in play.

Automatic recognition of ingestive-related behaviors of dairy cows based on triaxial acceleration

Ingestive-related behaviors including feeding and ruminating are important indexes to measure the health and welfare of dairy cows. The purpose of this study is to develop a method based on triaxial acceleration to automatically recognize feeding and ruminating of dairy cows. During the experiment, five diary cows raised in a barn were used as experimental subjects. A triaxial acceleration sensor was used as the device to collect jaw-movement data of dairy cows, and the behaviors of dairy cows were classified into three categories: feeding, ruminating and other behavior. The features of time-domain and frequency-domain were extracted from the raw acceleration data. Three machine learning algorithms including k-nearest neighbor, support vector machine and probabilistic neural network were used for the classification and the results based on four different data segment lengths were compared. The results show that the three algorithms can be used for recognition of feeding and ruminating with high accuracy. Under the condition that the sampling frequency of the sensor is 5 Hz, the combination of data segment length of 256 and k-nearest neighbor algorithm is the best scheme for recognition of feeding and ruminating in this study. The precision and recall of recognition for feeding were 92.8% and 95.6% respectively, and those of recognition for ruminating were 93.7% and 94.3% respectively. The specificity and AUC of recognition for feeding were 96.1% and 0.959 respectively, and those of recognition for ruminating were 97.5% and 0.959 respectively. This will provide an effective method for real-time monitoring of ingestive-related behaviors of dairy cows and lay a foundation for prediction of dairy cows’ health status and welfare to further achieve the purpose of disease prediction and adjusting feeding and management methods.

Kinematics and Statistics of Breaking Waves Observed Using SWIFT Buoys

Surface wave instrumentation floats with tracking were deployed by helicopter ahead of five large storms off the Oregon coast. The buoys drifted freely with the wave motions, surface currents, and wind. The buoys use a 9-DoF inertial measurement unit that fuses the measurements of accelerometers, magnetometers, and gyroscopes to measure acceleration in the global North-West-Up reference frame. Rapid sampling (25 Hz) allows for the observation of both propagating wave motions and wave breaking events. Bulk wave parameters and wave spectra are calculated from the motion of the buoys using conventional methods, and breaking wave impacts are identified in the raw acceleration data using a new algorithm based on a short-time Fourier transform. The number of breaking waves is used to infer breaker fraction, which is found to depend on bulk wave steepness as previously shown in the literature. The magnitude and duration of acceleration during breaking is used in a new quantification of breaker intensity, which increases with wave height, period, and steepness. There is significant variance of breaker intensity in a given wave field, such that intense breakers still occur in relatively mild wave fields. The buoy observations are compared to the output of the WaveWatch III forecast model, with evaluation of an empirical breaker prediction scheme applied to WaveWatch III output.

P.195Physical activity monitoring using wrist-worn accelerometer in the assessment of patients with myositis

This study aimed to evaluate physical activity (PA) monitoring using wrist-worn accelerometers in the assessment of myositis patients. Included patients had a diagnosis of dermatomyositis (DM), immune-mediated necrotizing myopathy (IMNM) or overlap myositis (OM) and were evaluated at baseline (M0) and 6 months (M6). IMACS core-set measures, ACR/EULAR improvement score, muscle endurance testing, 14-days raw acceleration data (expressed both as mean daily Euclidean norm minus 1 g (ENMO)), and quality of life questionnaires were recorded. Relationships between variables were investigated using Spearman correlation coefficient (ρ) and random forest (RF) regressions. Fifty-five patients (16 OM, 27 IMNM and 12 DM) were included. At M0, ENMO mainly correlated with manual muscle testing 8 (MMT8, ρ=0.44, p=0.001), creatinine level (ρ=0.43, p=0.001), HAQ (ρ=−0.53, p<0.001), and SF-36-physical functioning (ρ=0.40, p=0.002). According to RF regressions, most important features associated with ENMO Z-scores were HAQ, SF-36-physical functioning/energy/pain and disease duration. At M6, ENMO changes mainly correlated with changes in depression score (ρ=−0.57, p=0.001), HAQ (ρ=−0.51, p<0.001), and SF-36-energy/fatigue (ρ=0.51, p<0.001). According to the RF regression, most important features associated with ENMO changes were change in MMT8, HAQ, SF-36-physical functioning, physician global assessment, and depression score. Changes in ENMO were correlated with ACR/EULAR improvement score (ρ=0.57, p<0.01) and all patients, except one, achieving a major improvement at M6 had a significant ENMO change. Yet, variable changes in ENMO were observed in patients with minimal to moderate improvement at M6. PA levels were correlated with muscle strength and function, while PA change was also associated with emotional change. Only patients with major ACR/EULAR improvements at follow-up were associated with consistent changes in PA.

Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device.

Wearable, multisensor, consumer devices that estimate sleep are now commonplace, but the algorithms used by these devices to score sleep are not open source, and the raw sensor data is rarely accessible for external use. As a result, these devices are limited in their usefulness for clinical and research applications, despite holding much promise. We used a mobile application of our own creation to collect raw acceleration data and heart rate from the Apple Watch worn by participants undergoing polysomnography, as well as during the ambulatory period preceding in lab testing. Using this data, we compared the contributions of multiple features (motion, local standard deviation in heart rate, and “clock proxy”) to performance across several classifiers. Best performance was achieved using neural nets, though the differences across classifiers were generally small. For sleep-wake classification, our method scored 90% of epochs correctly, with 59.6% of true wake epochs (specificity) and 93% of true sleep epochs (sensitivity) scored correctly. Accuracy for differentiating wake, NREM sleep, and REM sleep was approximately 72% when all features were used. We generalized our results by testing the models trained on Apple Watch data using data from the Multi-ethnic Study of Atherosclerosis (MESA), and found that we were able to predict sleep with performance comparable to testing on our own dataset. This study demonstrates, for the first time, the ability to analyze raw acceleration and heart rate data from a ubiquitous wearable device with accepted, disclosed mathematical methods to improve accuracy of sleep and sleep stage prediction.

Deep Neural Network-Based Gait Classification Using Wearable Inertial Sensor Data.

Human gait has been regarded as a useful behavioral biometric trait for personal identification and authentication. This study aimed to propose an effective approach for classifying gait, measured using wearable inertial sensors, based on neural networks. The 3-axis accelerometer and 3-axis gyroscope data were acquired at the posterior pelvis, both thighs, both shanks, and both feet while 29 semi-professional athletes, 19 participants with normal foot, and 21 patients with foot deformities walked on the 20-meter straight path. The classifier based on the gait parameters and fully connected neural network was developed by applying 4-fold cross-validation to 80% of the total samples. For the test set that consisted of the remaining 20% of the total samples, this classifier showed an accuracy of 93.02% in categorizing the athlete, normal foot, and deformed foot groups. Using the same model validation and evaluation method, up to 98.19% accuracy was achieved from the convolutional neural network-based classifier. This classifier was trained with the gait spectrograms obtained from the time-frequency domain analysis of the raw acceleration and angular velocity data. The classification based only on the pelvic spectrograms exhibited an accuracy of 94.25% even without requiring a time-consuming and resource-intensive process for feature engineering. The notable performance and practicality in gait classification achieved by this study suggest potential applicability of the proposed approaches in the field of biometrics.

Evaluating and Enhancing the Generalization Performance of Machine Learning Models for Physical Activity Intensity Prediction From Raw Acceleration Data.

To evaluate and enhance the generalization performance of machine learning physical activity intensity prediction models developed with raw acceleration data on populations monitored by different activity monitors. Five datasets from four studies, each containing only hip- or wrist-based raw acceleration data (two hip- and three wrist-based) were extracted. The five datasets were then used to develop and validate artificial neural networks (ANN) in three setups to classify activity intensity categories (sedentary behavior, light, and moderate-to-vigorous). To examine generalizability, the ANN models were developed using within dataset (leave-one-subject-out) cross validation, and then cross tested to other datasets with different accelerometers. To enhance the models' generalizability, a combination of four of the five datasets was used for training and the fifth dataset for validation. Finally, all the five datasets were merged to develop a single model that is generalizable across the datasets (50% of the subjects from each dataset for training, the remaining for validation). The datasets showed high performance in within dataset cross validation (accuracy 71.9-95.4%, Kappa K = 0.63-0.94). The performance of the within dataset validated models decreased when applied to datasets with different accelerometers (41.2-59.9%, K = 0.21-0.48). The trained models on merged datasets consisting hip and wrist data predicted the left-out dataset with acceptable performance (65.9-83.7%, K = 0.61-0.79). The model trained with all five datasets performed with acceptable performance across the datasets (80.4-90.7%, K = 0.68-0.89). Integrating heterogeneous datasets in training sets seems a viable approach for enhancing the generalization performance of the models. Instead, within dataset validation is not sufficient to understand the models' performance on other populations with different accelerometers.

Automatic Identification of Bridge Vortex-Induced Vibration Using Random Decrement Method

Vortex-induced vibration (VIV) has been occasionally observed on a few long-span steel box-girder suspension bridges. The underlying mechanism of VIV is very complicated and reliable theoretical methods for prediction of VIV have not been established yet. Structural health monitoring (SHM) technology can provide a large amount of data for further understanding of VIV. Automatic identification of VIV events from massive, continuous long-term monitoring data is a non-trivial task. In this study, a method based on the random decrement technique (RDT) is proposed to identify the VIV response automatically from the massive acceleration response without manual intervention. The raw acceleration data is first processed by RDT and it is found that the RDT-processed data show different characteristics for the VIV response and conventional random response. A threshold based on the coefficient of variation (COV) of peak values of processed data is defined to distinguish between the two kinds of responses. Both random vibration and VIV for a three-DOF (degree-of-freedom) mass-spring-damper system are obtained by numerical simulation to verify the proposed method. The method is finally applied to the Xihoumen suspension bridge for identifying VIV response from three-month monitoring data. It is shown that the proposed method performs comparably with the method of novelty detection. A total of 60 VIV events have been successfully identified. Vortex-induced vibrations for the second to ninth vertical modes with modal frequency within 0.1~0.5 Hz occurs at wind velocity 5–18 m/s, with wind direction nearly perpendicular to bridge axis. Amplitude of VIV generally decreases with increase of wind turbulence intensity; however, noticeable VIV amplitude are still observed for turbulence intensity up to 13% in some cases.

Vibration‐based structural state identification by a 1‐dimensional convolutional neural network

AbstractDeep learning has ushered in many breakthroughs in vision‐based detection via convolutional neural networks (CNNs), but the vibration‐based structural damage detection by CNN remains being refined. Thus, this study proposes a simple one‐dimensional CNN that detects tiny local structural stiffness and mass changes, and validates the proposed CNN on actual structures. Three independent acceleration databases are established based on a T‐shaped steel beam, a short steel girder bridge (in test field), and a long steel girder bridge (in service). The raw acceleration data are not pre‐processed and are directly used as the training and validation data. The well‐trained CNN almost perfectly identifies the locations of small local changes in the structural mass and stiffness, demonstrating the high sensitivity of the proposed simple CNN to tiny structural state changes in actual structures. The convolutional kernels and outputs of the convolutional and max pooling layers are visualized and discussed as well.

0326 Sleep Stage Prediction With Raw Acceleration And Photoplethysmography Heart Rate Data Derived From A Consumer Wearable Device

0326 Sleep Stage Prediction With Raw Acceleration And Photoplethysmography Heart Rate Data Derived From A Consumer Wearable Device

Applying singular value decomposition on accelerometer data for 1D convolutional neural network based fall detection

The usefulness of applying singular value decomposition (SVD) on triaxial accelerometer data for one-dimensional (1D) convolutional neural network (CNN) based fall and activity recognition is investigated. Three-dimensional reduction methods, namely, SVD, sparse principal component analysis, and kernel principal component analysis, are compared for their effectiveness in extracting useful features for fall and activity recognition. Experiments conducted on three public falls and activities of daily living datasets show that SVD applied to acceleration data coupled with raw acceleration data or acceleration signal magnitude vector exhibited better 1D CNN fall and activity recognition accuracy than those using other principal component analysis based acceleration features.